Claude Code Team Explains Effort: Hardware Task Pass Rates Jump from 34% to 75%
AI寒武纪 · wechat · 2026-09-26
Thariq from the Claude Code team breaks down the effort parameter: it's a compute budget, and higher settings mostly buy deeper boundary-condition hunting and self-verification—not fixes for wrong initial approaches. On Terminal-Bench 3.0, Fable 5.1 going from lowest to highest effort cut test-omission bugs from 40 to 14, but requirement-misunderstanding failures rose from 25 to 47.
Gains vary hugely by domain: hardware pass rates jumped 34%→75%, security 64%→87%, ML 54%→73%, while routine software only 43%→56%. He recommends a four-step loop: have the model interview you into a spec, draft at low effort, iterate fast at low effort, then run final verification at high effort. Guidance: low for human-in-the-loop brainstorming and prototypes, medium for regular development, high for gnarly bug hunts, max for end-to-end complex builds. In Claude Code you can switch mid-conversation with /effort, and changing effort doesn't break prompt caching.
Related event: Anthropic Staff Explains How to Choose Claude Code's Effort Setting(5 posts)→
More from coding & agent
- FluentCRM MCP Server lets you manage marketing automation from Cursor or Claude — modelcontextprotocol · 2026-11-11
- Elvis Saravia: agents will subsume traditional software — and can even cut code bloat — omarsar0 · 2026-11-11
- Using a third-party inference provider with your coding harness hands it shell access to your machine — TheZachMueller · 2026-11-11
- Splash One: flight combat sandbox built with Opus 5.5 in PlayCanvas — willeastcott · 2026-11-11
- Next.js 16.5 delivers 3.3x faster builds for a leading AI lab — cramforce · 2026-11-11
- Plurai launches free agent simulation testing inside Claude Code and Cursor — PrajwalTomar_ · 2026-11-11